It is possible to exploit the panel design features of the CPS sample (see Chapter 4) to conduct longitudinal analyses with successive annual supplements. Half of the addresses that are included in the CPS ASEC sample in a given year were included in the sample the prior year. Such analyses encounter serious obstacles, however. The sample units are addresses, not the persons living at those locations. Persons who move during the year (about 14 percent of the population, based on recent estimates) would be excluded from any longitudinal analysis, introducing an obvious bias, as moving may be related to changes in circumstances that are relevant to medical care expenditures, resources, or family composition. Nonresponse to the supplement (about 15 percent currently) may introduce further bias in addition to reducing the number of sample households present in 2 consecutive years. On top of these concerns, the CPS is weaker than the other three surveys in its collection of data elements needed to model medical care expenditure risk.
In summary, none of the surveys is nearly as strong as we would like in its measurement of key baseline characteristics. With its strong measures of chronic conditions and very high-quality expenditure data, MEPS is clearly superior to SIPP. The HRS could provide a supplemental data source for the one-fifth of households that fall into the HRS universe. Estimates from the HRS could be used to validate the model estimates from MEPS for this segment of the population (or perhaps just the elderly), although differences in the variables available to serve as predictors would have to be taken into account.
Data Sources for Production of a Measure
Once a model of MCER has been developed, the estimates could be used directly (in MEPS), or the predictive model could be applied to another data set that provides measures of the relevant baseline characteristics. The latter approach offers a way to make the measurement of MCER more timely and to extend the measure to a larger and possibly more representative sample.
Because longitudinal data would not be required but production would impose other requirements, the set of surveys that could potentially serve this purpose is not the same as the three evaluated above. We include MEPS—but the full-year consolidated file rather than the longitudinal file. The consolidated file has an annual reference period and combines two successive, overlapping panels. We do not include the HRS or SIPP. The HRS represents too little of the population to be used for production. Although there are a number of issues with the use of SIPP for this purpose (see Czajka, in Part III), the survey’s abutted panel design has the